1. Field of the Invention
The present invention relates to data access technology preferably applicable to a network system, and particularly to a NUMA (Non Uniform Memory Access model) system.
2. Description of the Related Art
FIG. 21 shows a block configuration diagram showing an example of the NUMA system. In this system, each of several node devices 1000a, . . . , 1000n is connected to the other with a bus 1040 so that they can mutually send and receive the required information between them. Each node device has a CPU 1010a ( . . . , 1010n), a local memory 1020a ( . . . , 1020n) and a control unit 1030a ( . . . , 1030n). In addition, as shown in FIG. 22, the local memories 1020a, . . . , 1020n of the node devices constitute a common memory. The CPU 1010a, . . . , 1010n of the node devices are configured so that they can access such common memory.
In the NUMA system as described above, when the CPU 1010a, . . . , 1010n in the node devices 1000a, . . . , 1000n access the common memory, the latency, which is the period from the start of reading to the completion of data readout, differs among them. A cache is incorporated in each of the CPU1010a, . . . , 1010n so that the number of accesses from a certain node device (hereinafter referred to as the self node device or local node) to the cache in the CPU of the other node device (hereinafter sometimes referred to as the remote node) can be reduced.
If no restriction is imposed to copying of the data stored in the common memory to the cache memory, however, the data copy having the same address in the common memory may be stored in the cache memories of several processors, which would result in frequent accesses between the node devices. This spoils the latency of the access to the cache memory or the local memory of the remote node, which leads to a poorer system performance.
Further, since the local memories constituting the common memory are mounted individually one by one for the node devices in this system, the latency difference between the local node and the remote node becomes larger. This also results in a lower system performance.